vskip-flexidepth-qwen-3 — the Qwen3 FlexiDepth skippers of the vSkipper paper (v1.7)

folder	base (frozen)	routed layers	training (alignment-only ste_hard)	probe skip	served decode skip at the knee	native GSM8K	served knee GSM8K vs upstream	mean E2E at the knee	contents
qwen3-4b/	Qwen/Qwen3-4B @ 1cfa9a72	18–35 of 36	penalty 1e-5/4 (coef 2.5e-6), step 20,000	0.256	0.119	see paper App. K	see paper App. K	no resolved gain	full export
qwen3-8b/ (served)	Qwen/Qwen3-8B @ d117af2f	18–35 of 36	1e-4/4 to step 7,500, then 4e-4/4 (coef 1e-4) to step 15,000	0.520	0.416	70.43 (base 80.97)	75.06 vs 80.59	−7.3 ± 0.9 %	full export + records
qwen3-8b-2e-4-step10000/	same	18–35 of 36	1e-4/4 to step 7,500, then 2e-4/4 (coef 5e-5) to step 10,000	0.445	0.344	75.21	76.42 vs 80.52	−3.6 ± 1.0 %	router delta + records
qwen3-8b-2e-4-step18750/	same	18–35 of 36	1e-4/4 to step 7,500, then 2e-4/4 (coef 5e-5) to step 18,750	0.485	0.350	78.09	80.21 vs 80.59	−3.5 ± 3.9 %	router delta + records

The paper's main section serves qwen3-8b/; Appendix K reports all three Qwen3-8B checkpoints. Probe skip = the 32-prompt routing probe's chat-template skip share at training time (probe_step<N>-chat.json); served decode skip = the share of routed-layer decisions the always-route arm took through the Project-Only path at the GSM8K knee (from the served cell's attestation). Native GSM8K = lm-eval composite exact match on the 1,319 held-out rows (native_lmeval_results.json); served knee GSM8K = the same scorer on the served hybrid at 0.95 × Q* (served_knee_lmeval_score.json); E2E = matched-work mean end-to-end latency change against same-ladder upstream SGLang, three paired repetitions. Thinking is disabled by default in every folder's chat template.

Full exports hold the frozen base shards plus the trained router/projector tensors, the checkpoint's own modeling_ddqwen3.py and the tokenizer. The delta-only folders hold the trained tensors alone (step<N>_router.pt, 415 MB) with the same config, model code and tokenizer files; training/qwen3-flexidepth/build_export.py in the vSkipper harness rebuilds the full folder from the public base plus the delta, bit for bit:

python training/qwen3-flexidepth/build_export.py --base <Qwen3-8B d117af2f snapshot> --delta step18750_router.pt --ref . --out <dir>

Training recipe, sealed input contracts and scripts: training/qwen3-flexidepth/ in the harness.